Skip to content

Latest commit

 

History

History
44 lines (30 loc) · 1.41 KB

Usage.md

File metadata and controls

44 lines (30 loc) · 1.41 KB
Raw

Usage

This file will showcase how to reproduce our work.

Requirement

python 3.6+ (tested in 3.7)
PyTorch 1.1 or higher (tested in 1.2 & 1.3)
CUDA 9.0 or higher (tested in 10.0)
Linux (tested in Ubuntu 18.04)
apex (see NVIDIA/apex)
DALI (see NVIDIA/DALI)

Training

Classification

You can use the following commands to train a classifcation network.
We are training on 8 Tesla V100 GPUs. If you have 4, change "--nproc_per_node=8" to "--nproc_per_node=4".
For more detail parameters, please see main.py and main_mobile.py files.

# For normal networks, like ResNet. 
python3 -m torch.distributed.launch --nproc_per_node=8 imagenet.py -a fp10_resnet50 --b 32

Detection

Our detection codes are based on mmdetection framework. Thanks to mmdetection. For more details, please see mmdetection github.

# For training.
./tools/dist_train.sh local_configs/{config_file_name}.py 8

# For testing
python3 tools/test.py local_configs/{config_file_name}.py 
work_dirs/{model_path}/epoch_24.pth --gpus 8 --out work_dirs/{save_path}/{results_name}.pkl --eval bbox

Other tools

We also provide a series of related tools, such as visualization, analysis, and parameters/flops counter.